在數位經濟驅動下，零售與服務業正經歷線上線下整合(O2O)的轉型，消費者的「主動評估」行為已成為影響實體客流的前哨戰。然而在機車維修產業中，因其具備典型的「信任財」特質與高度專業技術門檻，消費者常面臨資訊不對稱的困境；且現有的需求預測模型多依賴歷史交易數據，存在預測誤差與滯後性，難以捕捉市場的即時變動。因此，本研究旨在構建一個整合Google Maps多維度數位足跡與內部歷史數據的新客量預測模型，探討數位訊號如何作為實體進店行為的預測指標。 本研究以X企業全台170間授權通路為實證樣本，蒐集 2025年2月至11月之內部維修紀錄與Google Maps外部資料。研究方法上，導入RoBERTa深度學習模型進行自然語言處理，萃取評論文字之情感分數與三大核心主題(技術、價格、服務)；同時結合GIS技術運算實體地理競爭特徵，並設計四階段消融實驗，運用正則化線性模型(Ridge、Lasso)與 XGBoost、LightGBM、隨機森林等樹狀整合演算法共同建構並比較預測框架。 研究結果顯示，數位足跡的預測價值具明確的模型條件性：純歷史的正則化線性模型在絕對精度上最為穩健，而多維數位足跡則於具非線性建模能力的整合演算法中、以及小型店家與高競爭等特定情境下，提供顯著的條件性邊際修正。實證發現，數位訊號在「小型店家」群組中具備顯著的「信任補償」效果，能協助品牌聲量薄弱的業者降低消費者的知覺風險；在高度競爭環境下，數位口碑更是決定客流流向的關鍵因素。此外，研究證實了「技術品質」相關的數位訊號對信任財產業的預測價值優於傳統「功能品質」。然而，研究亦發現技術話題提及率與新客流量呈現負向相關，顯示過度專業或涉及維修糾紛的技術討論在數位評估階段可能產生知覺障礙。本研究不僅填補了信任財產業預測領域的空白，亦能協助業者優化數位經營策略與資源配置。;Driven by the digital economy， the retail and service industries are undergoing a transformation toward Online-to-Offline (O2O) integration， where consumers′ "active evaluation" behavior has become a critical precursor to physical foot traffic. In the motorcycle maintenance industry， however， consumers often face information asymmetry due to the sector′s inherent characteristics as a credence good and its high technical barriers. Furthermore， existing demand forecasting models rely heavily on historical transaction data， which often suffer from prediction errors and time lags， making it difficult to capture real-time market fluctuations. Therefore， this study aims to construct a new customer volume forecasting model that integrates multi-dimensional digital footprints from Google Maps with internal historical data， exploring how digital signals serve as leading indicators for physical store visits. Using 170 authorized dealers of Enterprise X across Taiwan as empirical samples， this research collected internal maintenance records and external Google Maps data from February to November 2025. Regarding methodology， the RoBERTa deep learning model was employed for Natural Language Processing (NLP) to extract sentiment scores and three core themes (Technology， Price， and Service) from review texts. Simultaneously， Geographic Information System (GIS) technology was utilized to calculate physical competitive features. A four-stage ablation study was designed， comparing regularized linear models (Ridge， Lasso) against tree-based ensemble algorithms (XGBoost， LightGBM， and Random Forest) to construct and benchmark the forecasting framework. The results reveal that the predictive value of digital footprints is model-conditional: regularized linear models trained on historical data alone achieve the most robust absolute accuracy， while multi-dimensional digital footprints provide a significant conditional marginal correction within nonlinear ensemble algorithms and in specific contexts such as small-scale stores and high-competition environments. Empirically， digital signals exhibit a significant "trust compensation" effect for "small-scale stores，" helping businesses with weaker brand presence reduce consumers′ perceived risk. In highly competitive environments， digital word-of-mouth (eWOM) is a decisive factor in directing customer flow. Furthermore， the study confirms that digital signals related to "Technical Quality" possess higher predictive value for the credence goods industry than traditional "Functional Quality”. However， the study finds that the technical topic ratio is negatively correlated with new customer volume， suggesting that excessive technical details may increase perceived risk during the digital evaluation phase. This research not only fills a gap in the forecasting literature for the credence goods industry but also provides practical insights for operators to optimize digital marketing strategies and resource allocation.
